Conversational AI in HRIS: Self-Service in Pakistan

Conversational AI in an HRIS lets an employee ask a question in plain language and get an answer from the system of record: how many casual leaves are left, why this month's net pay changed, what the probation policy says. Done well, it answers in seconds and sends anything sensitive to a person. Done badly, it invents an answer about someone's salary.

The difference comes down to design rules. This guide gives you those rules for a Pakistani company, along with the questions employees ask most and the ones a bot should never answer alone.

In most companies a small set of questions produces most of the volume: leave balance, payslip details, policy rules, letters and certificates, and what happens next in a process. They share a feature. The answer already sits in a record or a document, so the bot looks it up and does not need to think.

Sort every question by risk before you automate it. The table gives a starting point.

Because the questions already arrive, just not through your HRIS. They come through WhatsApp, in person, by phone and by email, and each one interrupts someone. Our notes on leave approvals that cost more than they seem and on the employee who touches five systems a day describe where the time goes.

The adoption data shows HR leaders are moving. Gartner's 2024 survey of 179 HR leaders found 38% were piloting, planning or had implemented generative AI, up from 19% in June 2023. Dion Love, Vice President of Advisory in the Gartner HR practice, said more organisations were moving from exploring how generative AI might be used to implementing solutions (Gartner, February 2024). That count blends three stages, so it does not mean 38% have it running. For you, the message is that employees will soon expect the answer in a chat window, because they get one everywhere else.

Five rules do most of the work. Answer from records and approved documents, not from the model's memory. Check who is asking before showing personal data, so a manager sees team figures and an employee sees only their own. Show the source of every answer. Say "I do not know" and hand off when the source is missing. Log every conversation that touches personal data.

I run HR at Bitsbuffer, and my read is that the third rule matters most. An answer with a visible source, such as the policy section and its date, lets an employee check it in ten seconds. An answer without one turns every conversation into an argument about what the bot meant.

Take two questions. An employee asks how many casual leaves remain. A good answer gives the number, the date it was calculated and a link to the leave ledger: 4 casual leaves remain as of today, with 2 pending approval. It reads the value from the record. It does not estimate it.

A second employee asks whether probation can be extended. A good answer quotes the section of the policy, shows its version date and says who decides: the manager recommends and HR confirms. Then it offers to start the request. If the policy is silent, the right answer is that the policy does not say and that HR will reply within a stated time.

Both answers do three things. They cite a source, they stay inside what the record says and they leave a trail. If a vendor's demo produces smooth answers with no source shown, ask where each one came from.

Ask whether the assistant understands English, Urdu and Roman Urdu, because many employees will write the way they text. Test it with real phrasing from your own team, such as a leave question written in Roman Urdu, before you sign anything.

Decide the channel on purpose. A chat inside the employee portal is easy to secure. WhatsApp is where people already are, but it needs strict identity checks and limits on what the bot may show. For personal data, restrict access by role and keep a retention rule. DLA Piper's Pakistan guide describes the Personal Data Protection Bill as a draft not yet promulgated, so check the current status with a lawyer and set your own limits meanwhile (DLA Piper). Our guide to ethical AI governance in an HRIS lists the controls.

Track four numbers each month. Resolution rate: the share of questions answered without a person. Escalation rate: how often it handed off, which should be higher at first than you expect. Accuracy: HR checks a random sample of 20 conversations every week against the records. Employee rating: a one-tap thumbs up or down.

Start with the three highest-volume, lowest-risk questions, which are usually leave balance, holiday calendar and policy lookups. Add payslip questions only after the sample checks stay clean for a full month. Workflow Engine is Bitsbuffer's HRMS for Pakistani companies, and we suggest holding it to this list. Bring your top ten employee questions to a demo and we will say plainly which ones it answers today and which it does not.

It should never decide a leave dispute, explain away a payroll error, handle a complaint about a colleague or a manager, or advise on someone's legal rights. It also cannot fix a policy nobody can explain. If two managers read your leave policy differently, the bot will repeat the confusion faster. Fix the policy first.